How eight sessions over seven weeks converged on one way to run an agent org — and the gotchas each one paid for.
The operating style was never designed in one sitting — it converged. Across ~8 sessions from mid-May to late-June 2026, each paid for one more piece of the fix. They all point at four moves: Codex builds, the lead orchestrates, watchers heal, and the founder reviews outcomes — not internals.
Each session in this arc delegated a build to a fleet, hit a wall, and bolted on one piece of the fix. Read top-to-bottom, the column on the right is the operating style assembling itself.
| Date | Session | What it contributed |
|---|---|---|
| 05-11 | Paperclip postmortem → Runfusion design | First hard postmortem of the multi-agent build experiment + a next-gen conductor design |
| 05-31 | Assessment of Agentic Work Control | Named the core gap — the autonomous loop existed as a goal, but live work still leaned on manual coordination across chats, branches, PRs, Linear, Beads, RunFusion |
| 06-01 | PKM Watchdog dispatch + remediation | The self-healing-watcher lineage: codex-driven watchdog, health checks, failure-aware notifications |
| 06-02 | Multi-agent Workflow gotchas (learning) | The Claude+Codex split with Codex as second-opinion; the non-obvious gotchas |
| 06-07 | Safe autonomous agent-lane orchestration (learning) | The release train; the surface-only-PR failure → contract propagation, cross-engine review, fail-closed-merge, reversibility-replaces-review |
| 06-08 | Codex-heavy 80/20 salvage (learning) | The model-routing style: Codex carries the token load, Claude only orchestrates |
| 06-29 | PKM Orchestration deep dive | This build's immediate predecessor — pivoted to Paperclip/Hermes as delegated, Codex-governed agents |
| 06-29 | Hermes↔Paperclip bridge + Donna-autonomous | Made Donna act on Paperclip work; gpt-5.5 / openai-codex for the bridge, explicitly not Claude |
| 06-29 | This session — factory consolidation | COO anti-illusion 2IC, dept-heads-own-goals, deterministic dispatch, canon-lock, watchdog auto-heal |
The shape repeats every time: a human delegates a build to a fleet, the early attempts produce surface-only or illusion output, the session adds one piece of the fix, and the style converges. By this session, the four moves are stable enough to name.
gpt-5.5 in worktrees; 5.4-mini for light work; never spark. Opus never burns tokens on the build.
Scopes, verifies, surfaces, owns the contract. Coordinates the fleet — doesn't do its work.
Git-init, relaunch, auto-resume, escalate. Detection without action is a non-fix.
URLs and decisions, not paths and PRs. Reversible work never waits for approval.
A loop fired on schedule, spent the tokens, and shipped nothing — every dashboard green.
The reusable form — landed as a docs/solutions learning in the PKM monorepo:
heartbeat:invoke. Round-robin real lanes across team goals with caps + an idempotent lock, instead of a no-op wake that only looks like work.The recurring failure modes — the "never repeat" reference set. Each is a tax a past session already paid so this one didn't have to.
codex-spark leaked into launchers, exhausted quota, took down a CEO agent. → 5.5 / 5.4-mini, never spark.ssh exit 255pkill. → prewarm ControlMaster, base64-pipe, kill by PID.printf parse errorprintf of prose.--company-id rejectedainativeplatform = OLD. → verify repo identity against canon."Running" is proven by output, not uptime.
Codex builds; the lead judges.
Detect-and-heal, never detect-and-report.
Canon-lock kills drift before it spends.
The founder reads URLs, not internals.
"Done" means Landed — or it isn't done.
"Running" is proven by output, not uptime. Every other pattern in this lineage is a way of making that true.